Benchmark guided work selection

[retina]

Originally Posted by S00113

...
Best time for 61 bit trial factors: 19.153 ms.
Best time for 62 bit trial factors: 39.511 ms.
Best time for 63 bit trial factors: 39.585 ms.
Best time for 64 bit trial factors: 22.897 ms.
...

Seems to me that the code should be set to use the 64bit+ algorithm for 62 and 63bit factoring!

[henryzz]

Originally Posted by retina

Seems to me that the code should be set to use the 64bit+ algorithm for 62 and 63bit factoring!

yes
as far as i am aware a similar thing happens on all 32-bit pcs

[S00113]

Originally Posted by henryzz

yes
as far as i am aware a similar thing happens on all 32-bit pcs

I greped for "to 2\^63" inn all results.txt files from all of my 102 AMD64 64bit cores doing normal TF, and got no results. None ever for a total of 144218 lines matching "to 2^". If this is a general problem for 32 bit, then why isn't all this work sent to 64bit cores, letting 32bit cores do the other bit levels? Count this as another example of how benchmark guided work selection could speed up GIMPS a lot.

[garo]

I think the reason for no match is that normal TF usually begins with factoring to 2^63 already done by LMH.

[S00113]

Originally Posted by henryzz

yes
as far as i am aware a similar thing happens on all 32-bit pcs

I did a few more benchmarks on different 32 bit CPUs, and this is not true.
Here is a Pentium II 350 MHz:

Best time for 58 bit trial factors: 41.833 ms.
Best time for 59 bit trial factors: 41.897 ms.
Best time for 60 bit trial factors: 41.744 ms.
Best time for 61 bit trial factors: 41.503 ms.
Best time for 62 bit trial factors: 77.971 ms.
Best time for 63 bit trial factors: 78.024 ms.
Best time for 64 bit trial factors: 172.369 ms.
Best time for 65 bit trial factors: 174.874 ms.
Best time for 66 bit trial factors: 177.939 ms.
Best time for 67 bit trial factors: 178.088 ms.

Almost half the speed at 62 and 63 bit, and less than quarter speed above 64 bit. For the Atom the speed of 64+ bit was roughly the same as up to 61 bit.
Here is a Pentium III 1100 MHz:

Best time for 58 bit trial factors: 13.794 ms.
Best time for 59 bit trial factors: 14.130 ms.
Best time for 60 bit trial factors: 14.339 ms.
Best time for 61 bit trial factors: 13.868 ms.
Best time for 62 bit trial factors: 24.615 ms.
Best time for 63 bit trial factors: 24.714 ms.
Best time for 64 bit trial factors: 55.378 ms.
Best time for 65 bit trial factors: 56.591 ms.
Best time for 66 bit trial factors: 57.133 ms.
Best time for 67 bit trial factors: 57.328 ms.

More or less the same as Pentium II, just a little less speed drop above 64 bit and of course much faster over all. Notice that the Pentium III is much faster than the Intel Atom up to 63 bits, even with a lower clock frequenzy. For 64 bits and above it is the other way around.

An Intel Xeon 2.7 GHz:

Best time for 58 bit trial factors: 10.700 ms.
Best time for 59 bit trial factors: 10.691 ms.
Best time for 60 bit trial factors: 10.755 ms.
Best time for 61 bit trial factors: 10.722 ms.
Best time for 62 bit trial factors: 10.723 ms.
Best time for 63 bit trial factors: 11.691 ms.
Best time for 64 bit trial factors: 11.729 ms.
Best time for 65 bit trial factors: 14.411 ms.
Best time for 66 bit trial factors: 14.397 ms.
Best time for 67 bit trial factors: 14.386 ms.

This one shows less than 10% slowdown at 62 and 63 bits, and only 40% slowdown at 65 bits and above.

If theese four were set to factoring, there would be a significant increase of throughput if the PII and PIII got 58 to 62 bit factoring, the Intel Xeon got 63 and 64 bits, and the Atom got 65 bits and above, compared to random work assignment. (Of course one have to make adjustments based on more benchmarks and which work is queueing up on the server.)

[garo]

Thanks for the great work Sturle. I've noticed this too and as a result have taken to running benchmarks on my machines to figure out the most efficient TF bits. If course I have only 7 CPUs unlike your hundreds so it is an easy task for me. But you are right, it would be nice to use TF benchmarks to assign the best bit ranges to a computer.